Linking distinct spatial omics datasets

The spatial multi-omics analysis tool SpaMosaic presents a method for integrating datasets that share only some measured features and estimating missing ones. In a study published in Nature Genetics, the researchers assessed its performance in identifying spatial regions using simulations and experimental data from several tissues. The aim is to connect datasets collected with different technologies and under different conditions to build comprehensive tissue maps.

Spatial omics combines information about molecular features with their locations in tissue. When datasets differ in their measured modalities, such as RNA, proteins, or epigenomic features, it can be difficult to link them. This is why “mosaic integration,” which combines datasets that overlap in only some features, is needed. Correcting for differences caused by measurement batches is another challenge in integrated analysis.

Identifying spatial regions and estimating missing information

SpaMosaic uses contrastive learning and a graph neural network to construct a latent space that is independent of modality and corrected for batch effects. This space provides a shared representation of features across datasets and is used to identify spatial regions and estimate missing modalities.

The comparisons included simulated data and experimental data from brain, embryo, tonsil, and lymph node. The measured features covered RNA and protein abundance, chromatin accessibility, and histone modifications. This describes the scope of the datasets included in the evaluation; it does not mean that every feature was measured together in every tissue.

The researchers reported that SpaMosaic outperformed existing integration methods in identifying consistent spatial regions while reducing noise and batch effects. It also distinguished fine anatomical structures in real datasets that differed in technology, developmental stage, spatial resolution, and modality composition, and was used to build a comprehensive spatial map of a mouse embryo. The available abstract does not name the comparison methods or report the percentage improvement in performance.

In mosaic data from mouse brain, the method estimated missing histone modifications. The researchers said the estimates reproduced expected correlations between transcriptomic and epigenomic features and captured more region-specific regulatory associations than measured chromatin-accessibility data. These findings suggest that relationships can be explored even in datasets where different modalities were not measured together. However, estimated regulatory associations should be distinguished from proof of a causal mechanism.

Potential for large-scale tissue maps and validation needs

The researchers also assessed scale. They reported that SpaMosaic could integrate more than 100 sections and, separately, process a single section containing more than 800,000 spots. Spots are spatial measurement locations; this figure should not be interpreted as a cell count. The abstract does not provide processing time, memory use, or required hardware.

SpaMosaic offers an analytical framework for linking heterogeneous spatial omics datasets to explore tissue structure and molecular features together. Estimates for missing features could help identify candidate regulatory associations to investigate in follow-up experiments. Validating these estimates remains a practical challenge. The study evaluated the computational method’s performance and applications; it did not test improvements in diagnosis or therapeutic effects.

The methods and main findings are available in the paper abstract indexed by Europe PMC.